Monitoring Intersection Safety Using Infrastructure LiDAR

Summary & Methodology Analysis

Urban intersections are hazardous locations requiring real time monitoring to anticipate conflicts, but deploying surrogate safety measures at the edge is challenging due to the need for site specific trajectory data, noisy raw perception outputs, and strict requirements for deterministic real time execution. To address this, the PRISA framework operates through a sensing and perception layer that uses LiDAR sensors to provide continuous 3D point cloud coverage, applying 3D object detection and multi-object tracking to output bounding box parameters and semantic class labels. The pipeline then splits into a self-supervised data curation phase and a real time trajectory prediction phase. Phase I accumulates perception outputs over extended periods, filters out incomplete or noisy tracks, interpolates brief gaps, downsamples stationary objects to create a training database without manual annotation, and trains a trajectory prediction model, which forecasts the path of a moving entity, until convergence. Phase II applies speed-based filtering to partition objects into dynamic vehicles, slow or stationary vehicles, and vulnerable road users, uses a prediction model for dynamic vehicles and vulnerable road users while holding stationary vehicles constant, and extracts observation sequences to forecast future trajectories. Finally, surrogate safety assessment identifies potential conflict partners within a velocity-dependent region of interest, classifies pairwise interactions into longitudinal, crossing, and vulnerable road user involved categories using heading alignment and spatial position, and evaluates conflict risk using Time-to-Collision for longitudinal conflicts and Predicted Post-Encroachment Time for crossing and vulnerable road user interactions.

The framework integrates several models and datasets, including R-LiViT, FlowChain, MID, Trajectron++, Constant Velocity, Constant Acceleration, PointPillars, and AB3DMOT. Among these, FlowChain achieved the lowest Average Displacement Error and Final Displacement Error of 0.646 m / 1.111 m on R-LiViT Int.2, 0.491 m / 0.710 m on R-LiViT Int.3, and 0.505 m / 0.928 m on the Georgia M.L.K. field site at 5 Hz. These evaluations demonstrate how the system handles trajectory forecasting under real world deployment conditions across different field sites.

Despite its performance, PRISA faces specific limitations. The trajectory predictor may produce a straight-motion bias or turning trajectories that overlap with stopped vehicles or roadside objects, yielding artificially low Predicted Post-Encroachment Time values despite limited actual risk. Additionally, the R-LiViT dataset's fragmented structure of short sequences is structurally incompatible with Phase I self-supervised training which requires continuous long-duration streams, necessitating the use of ground-truth annotations as inputs.